How2Sign

该数据集是一个多模态的开词汇、带字幕的美国手语视频集，时长约80小时，内容为教学视频的美国手语翻译，覆盖了广泛的主题。这些视频与其时间对齐的字幕一起使用，用于训练和评估检索模型。数据规模包括31,075个训练样本、1,739个验证样本以及2,348个测试样本的视频-字幕对。该数据集的任务是使用自由格式的文本查询进行手语检索。

This dataset is a multimodal open-vocabulary, subtitled American Sign Language (ASL) video corpus with a total duration of approximately 80 hours. It consists of ASL translations of educational videos covering a wide range of topics. These videos, paired with their temporally aligned subtitles, are used for training and evaluating retrieval models. The dataset comprises 31,075 training video-caption pairs, 1,739 validation video-caption pairs, and 2,348 test video-caption pairs. The task supported by this dataset is sign language retrieval using free-form text queries.